high performance computing on graphics processing units: hgpu.org

Recent graphics processing units (GPUs) integrate wimpy microcontrollers on a chip. They are often used to execute firmware code configuring the functional units of GPUs. This paper opens up the programming of these microcontrollers and explores how to utilize them for GPU resource management. Our prototype system provides a compiler suite for NVIDIA’s GPU microcontrollers with its basis on the Low Level Virtual Machine (LLVM) infrastructure. As a proof of concept, we develop fully-functional firmware using our compiler and provide a basic performance evaluation. The experimental results demonstrate that the overhead of introducing our firmware is suppressed to within 2.3%, as compared to the native proprietary firmware, while the impact of overhead is no greater than 0.01% of the total execution time according to microbenchmarks.